Network traffic dynamics prediction with a hybrid approach: Autoencoder-VaR

Network-wide traffic prediction is more effective for implementing traffic management control than traffic prediction for a single road. In order to improve the efficiency of network traffic forecasting, this research proposes a hybrid machine learning-based model, the Autoencoder-VAR (AE-VAR), which takes traffic time series from all locations of interest as input and performs predictions for network-wide locations simultaneously. Firstly, the Autoencoder is used to extract the essential features of the original data, retain the spatial-temporal dynamic effects between traffic flows and exclude random noises as much as possible. Then, the extracted feature time series are modeled and predicted with a VAR model at a lower dimension. Finally, the predicted features are projected back to the original data space. This methodology can take into account interactive dynamics of traffic flows between adjacent roads within the entire network with a less complicated model structure than many existing models. The empirical study on an urban road network using ground truth data indicates that the proposed AE-VAR model can effectively improve the accuracy of traffic predictions at network level. The proposed model structure is an efficient approach for network-scale traffic prediction.